Alibaba has launched Qwen3, the latest addition to its open-source large language model (LLM) family, delivering a major leap in AI capabilities with hybrid reasoning, extensive multilingual support, and superior agent-based functionality — all designed to fuel next-generation applications from smart glasses to autonomous vehicles.

The Qwen3 series, unveiled globally, introduces six dense models and two Mixture-of-Experts (MoE) models, reinforcing Alibaba’s commitment to open access AI innovation. Available on platforms such as Hugging Face, GitHub and ModelScope, the models target developers eager to build AI-powered tools across a wide array of devices and industries.

Qwen3 also debuts Alibaba’s first hybrid reasoning architecture, allowing models to fluidly switch between “thinking” and “non-thinking” modes. The thinking mode is tailored for tasks demanding advanced reasoning, such as mathematics and coding, while non-thinking mode delivers rapid, general-purpose responses.

Developers tapping into Qwen3 via API can finely adjust the model’s “thinking duration”, supporting up to 38,000 tokens. This gives them control over performance and computational efficiency. Notably, the Qwen3-235B-A22B MoE model promises lower deployment costs compared to rival models, positioning it as a cost-effective yet high-performing option for AI builders.

Trained on an immense 36 trillion tokens — double the dataset used for its predecessor Qwen2.5 — Qwen3 sets new benchmarks in reasoning, tool use, multilingual capabilities, and human alignment.

Among its standout features, Qwen3 supports 119 languages and dialects, boasting leading performance in translation and instruction-following. It also integrates seamlessly with agent-based applications through native support for the Model Context Protocol (MCP) and robust function-calling abilities.

The model has also demonstrated top-tier results across key industry benchmarks including AIME25 for mathematical reasoning, LiveCodeBench for coding, and Arena-Hard for instruction-following. Its development relied on a four-stage training approach involving chain-of-thought prompting, reinforcement learning, and fusion techniques to refine hybrid reasoning.

Qwen3 has already been adopted in Alibaba’s own AI-powered assistant app, Quark, and API access is on the horizon via its Model Studio platform.

Since the Qwen series debuted, it has become a force in the open-source AI space. With more than 300 million downloads worldwide and over 100,000 derivative models created by developers, Alibaba’s latest release looks set to cement its growing influence.